Reduce DAEFunction specialization latency

[ChrisRackauckas-Claude]

Important

Ignore this PR until it has been reviewed by @ChrisRackauckas.

What changed and why

Apply Base.@nospecializeinfer to the ModelingToolkit DAEFunction wrapper and avoid specializing its model-specific u0, p, analytic, and initialization_data keyword values. This matches the existing ODEFunction wrapper strategy and reduces inference work during standard DAE construction without changing the public API or runtime behavior.

Compile-latency evidence

I measured a three-state Robertson DAE with default initialization on Julia 1.12.7, one Julia thread, and a fresh Julia process for every sample. Package loading happened before the @timed DAEProblem(...) interval. Each constructed residual evaluated to [0.0, 0.0, 0.0].

Exact current-base comparison (median of 3)	Parent e91f13e48f	This PR f0a0b8aab2	Change
DAEProblem time	9.738 s	9.501 s	-2.44%
compilation time	9.679 s	9.442 s	-2.45%
allocations	844,725,672 bytes	843,232,560 bytes	-0.18%

The individual total-time samples were 9.436, 9.799, 9.738 seconds on the parent and 9.272, 9.726, 9.501 seconds on this PR.

An earlier isolated-depot run on the immediately preceding upstream commit showed a larger effect: median constructor time 11.696 -> 9.833 seconds and compilation time 11.616 -> 9.774 seconds (-15.9% for both). Trace-compile timing in that environment reduced the typed DAEFunction constructor from 3.49 s to 1.76 s and its model-specific keyword wrapper from 2.12 s to 1.57 s. The exact-current-base result above is the conservative headline measurement.

The benchmark script was temporary and was removed after collecting the samples, as required for local development scripts.

Verification

GROUP=InterfaceI julia +1.12 --startup-file=no --project=lib/ModelingToolkitBase -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Broken  Total      Time
InterfaceI    | 1601       5   1606  42m05.5s
Testing ModelingToolkitBase tests passed

julia --startup-file=no --project=@runic -m Runic --check --diff lib/ModelingToolkitBase/src/problems/daeproblem.jl
git diff -U0 -- lib/ModelingToolkitBase/src/problems/daeproblem.jl | typos -
git diff --check

All three commands exited 0 with no output.

GROUP=QA was also run. It reported 20 passing checks and 1 failing package-wide JET check. The identical command reproduced the same failure on a detached clean upstream/master at e91f13e48f: JET returned 450 reports. A parallel investigation bisected the QA-lane boundary to 92c27bb and added current reproduction evidence to #4958 (comment). No test was skipped, disabled, allowlisted, or silenced.

CI audit

The full PR matrix reached a terminal state with 97 passed checks, 7 failed checks, 2 canceled checks, and 1 skipped check.

• The package-wide QA JET failure reproduced identically on the exact upstream parent and is tracked at QA: JET package analysis exceeds the 60-minute root QA budget #4958.
• Four Optimization failures reproduced on the exact upstream parent and were reduced to a regression in CasADi.jl 1.3.0. The fix is merged at Restore SymbolicUtils code generation support CasADi.jl#42; the 1.3.1 release PR is green at Release v1.3.1 CasADi.jl#44.
• The downgraded QA failure reproduced on the exact upstream parent. It comes from test extras promoted into runtime dependencies by julia-downgrade-compat; the separate workflow fix is green at Restore sublibrary project metadata before downgrade QA .github#131.
• The docs failure reproduced in a complete clean-parent docs build and is an existing stale-link failure. Its focused fix has a green docs job at docs: replace stale Pantelides link #5056.
• Two Optimization Julia LTS jobs reached their two-hour job limit without a test error annotation. One exact-parent counterpart also timed out; the other exact-parent counterpart passed. GitHub did not permit targeted reruns of either matrix job, so the second cancellation remains inconclusive rather than established as an upstream baseline.

No behavior-specific test was added because this change consists only of compiler specialization annotations; the existing InterfaceI coverage exercises DAE construction, while the fresh-process benchmark is the discriminating performance check.

Not verified locally

• GROUP=Everything, GPU, and downstream groups were not run locally.
• The docs build was not run locally for this branch because it changes no public API, docstring, or documentation file.

🤖 Generated with Codex CLI 0.151.0 (model: gpt-5.6-sol; session: 01a059b1-5fa2-76a2-937a-33502aa35de5).